Abstract
Reconstructing three-dimensional volumetric patterns from sparse, near-field two-dimensional projections remains a fundamental challenge in computational imaging and multi-modal perception, where depth ambiguity and occlusion-induced phantom artefacts routinely undermine reconstruction fidelity. To address this ill-posed inverse problem, we propose a Hierarchical Physics-Constrained Information Fusion architecture that progressively collapses spatial uncertainty through sequential physical, statistical, and geometric constraints. Rather than relying on late-stage heuristic overlay, the proposed architecture operates across three inference tiers. In Tier 1, the Physical Constraint Prior, a multi-channel convolutional neural network surrogate acts as a physics-informed encoder, mapping discrete hardware parameters to spatial tensors and circumventing computationally prohibitive Monte Carlo simulations while tightly bounding the physical solution space. In Tier 2, the Statistical Uncertainty Estimation, a dynamic-response generalized expectation-maximization algorithm synthesizes distance-correct response matrices, adaptively resolving near-field depth mismatch and restoring focal sharpness with provable monotonic log-likelihood ascent. In Tier 3, the Geometric Uncertainty Collapse, a distributed Logarithmic Opinion Pool consensus protocol cross-validates two-dimensional probability heatmaps against three-dimensional exteroceptive structural constraints, enforcing Boolean intersection across orthogonal views to deterministically eliminate phantom artefacts. Extensive physical experiments in heavily occluded environments demonstrate that the proposed paradigm reduces the mean three-dimensional localization error from 0.86 ± 0.29 metres to 0.05 ± 0.01 metres. This architecture establishes a generalizable probabilistic framework for sparse-view tomographic pattern recognition, bridging physics-enhanced learning and multi-agent geometric consensus.